On October 7, Anthropic released Claude Haiku 5.5, an updated version of its "small" model that, by the company's own measurements, costs several times less than its predecessor: input tokens now start at $0.10 per million, roughly a 75% drop from Haiku 4.5. The next day, October 8, Anthropic added two new tools to Claude: Dashboards, which builds data dashboards from natural-language instructions and connects to Snowflake, Databricks, and Salesforce, and Motion, which generates editable animated videos from a text description without using video-generation models that depict real people.
Why a small business even needs a "small" model
At first glance, a cheaper model sounds like a technical detail of interest only to developers. In practice, it solves a very concrete problem that almost every business runs into when trying to roll out AI systematically rather than piecemeal: expensive models make sense for complex tasks, but stop being worthwhile once you need to process thousands of repetitive operations — sorting incoming requests, tagging CRM contacts, trimming product descriptions for a catalog, transcribing and structuring call-center calls.
Picture a small chain of three dental clinics that wants to automatically classify incoming WhatsApp and email messages as a complaint, an appointment request, a billing question, or spam. At a volume of 2,000–3,000 messages a month, the gap between a model priced at $3 per million tokens and one priced at $0.10 is the difference between a project that pays for itself in a quarter and a project that stays forever in the "nice idea, too expensive" pile. Haiku 5.5 moves exactly this kind of task from "too expensive" to "no longer too expensive."
Dashboards deserves a separate mention. For a business without a dedicated data analyst, it's a way to get a working sales or lead dashboard in an hour instead of weeks of back-and-forth with a contractor, with automatic updates as new data comes in and visible logic behind every chart — which, being a beta, obviously still needs regular manual checking rather than blind trust in the numbers.
What's worth reading with some skepticism here
All the performance figures — 72.4% on the OSWorld 2.1 benchmark versus 15.7% for Haiku 4.5 and 48.9% for OpenAI's GPT-6 Luna — are Anthropic's own measurements, published on the company's own blog. That doesn't make them wrong, but it does mean that before moving real workflows onto the new model, it's worth running your own actual tasks through it and comparing answer quality rather than relying solely on a marketing benchmark. Experience with deploying AI in business processes almost always shows the same pattern: a model that excels at a synthetic test can still stumble on the specifics of a particular industry — medical terminology, legal phrasing, the quirks of a regional market.
The second limitation is that Motion creates video through animation code rather than neural generation, which is both an advantage and a drawback: the advantage is no strange artifacts or "melting" faces; the drawback is that it's animation, not live footage, so for videos where a real person matters — a customer testimonial, for instance — the tool simply doesn't fit by its very nature.
A third point Anthropic doesn't address openly: cutting the price of small models isn't charity, it's a strategy to keep developers inside its own ecosystem at a time when, according to OpenRouter data, OpenAI overtook Anthropic in weekly model spend for the first time in two and a half years, in early October. Price competition between AI companies is good for the businesses that use these tools, but it also means today's prices are no guarantee of next year's prices.
For us at Dayava, this news is one more confirmation of something we already see in client work: good automation isn't about "putting the most powerful model everywhere" — it's about matching the right model to each task on a price-to-quality basis, with hands-on oversight at launch and gradual scaling after that.
If a similar solution for your business sounds interesting, leave a request at dayava.pt/contactos/.